# Does Humanoid Locomotion Have a Depth Sensor Problem Nobody's Solving Correctly?

**Yes — and a new paper from a team of eleven researchers argues the entire field has been approaching it wrong.** Published today on arXiv (2609.11553), CAP — Continuously Adaptive Perception-Blind Humanoid Locomotion via Learned Denoising — presents a single-stage locomotion policy for the [Unitree Robotics](https://humanoidintel.ai/companies/unitree-robotics) G1 that reconstructs usable depth information from corrupted sensor input, rather than abandoning it or routing around it. The core claim: when a depth sensor partially fails mid-deployment, existing perceptive policies break badly and binary-switching architectures throw away recoverable signal. CAP does neither, and the authors validate it in both indoor and outdoor trials on real hardware.

The practical stakes are significant. Depth sensors — structured light, stereo cameras, ToF arrays — are notoriously unreliable in uncontrolled environments. Sunlight washes out IR-based depth. Transparent surfaces produce phantom readings. Occlusion from the robot's own limbs creates intermittent blind spots. Any humanoid operating outside a warehouse will face these conditions routinely.

## The Fundamental Architecture Bet the Field Has Been Making

The dominant pattern in perceptive humanoid locomotion research has been clean-data assumption: train a policy on depth observations, assume those observations arrive intact, deploy. When that assumption breaks — and in real outdoor environments, it always does — performance degrades unpredictably.

A secondary pattern has emerged recently: routing or switching architectures that detect whether perception is trustworthy and hand off to a blind [proprioception](https://humanoidintel.ai/glossary/proprioception)-only sub-policy when it isn't. The CAP authors identify a structural flaw here: partially corrupted depth still contains recoverable geometric information. A hard switch to blind control discards that signal entirely, even when 60% of the depth image is still valid.

Both approaches fail at the same point — the transition between "perception working" and "perception dead." Real-world failure modes are rarely binary.

## What CAP Actually Does

The system is built around two components that operate simultaneously rather than in sequence:

**A perceptive world-model encoder trained as a learned denoiser.** Rather than passing raw depth to the policy, this encoder attempts to reconstruct clean depth from whatever corrupted signal is available. When depth is clean, it functions as a standard perceptive encoder. When depth is heavily corrupted, it extrapolates from available geometry. The policy never sees the raw corrupted input directly — it sees the encoder's best reconstruction attempt.

**A co-active proprioceptive variational encoder.** This runs in parallel, supplying body-state information that contains zero depth dependency. Crucially, this isn't a fallback — it runs continuously alongside the world-model encoder, providing the policy with ground-truth body state regardless of perceptual quality.

The training recipe is where the engineering insight lives. The authors couple two mechanisms: a depth-noise curriculum applied to the world-model encoder input, and world-model feature dropout applied to the policy-facing latent. This forces the policy to train across the entire perception-quality spectrum — from perfect depth to total depth failure — in a single end-to-end pass. The policy learns to weight encoder outputs appropriately based on implicit confidence signals baked into the latent representation.

This is a materially different bet from the switching architectures. There's no explicit "am I perceiving correctly?" classifier. The denoising encoder implicitly communicates confidence through the structure of its latent output, and the policy learns to read that signal.

## Simulation and Hardware Results

In simulation, the authors report that CAP matches or improves upon perceptive baselines when depth remains informative — meaning the denoising step doesn't introduce a performance tax under clean conditions. More importantly, performance degrades more smoothly than a binary-switching baseline as perception worsens. The degradation curve matters as much as the floor: a policy that fails gracefully is operationally safer than one that works perfectly until a threshold and then collapses.

On the Unitree G1, the team ran controlled trials and open-ended indoor-outdoor deployments under three real-world failure modes: intermittent occlusion, real-sensor corruption, and outdoor depth artifacts from direct sunlight. The paper describes the results as demonstrating "perception-robust locomotion" across all three conditions, though specific quantitative metrics from the hardware trials would need to be drawn from the full paper.

The [sim-to-real transfer](https://humanoidintel.ai/glossary/sim-to-real-transfer) story here is non-trivial. Noise curriculum during training appears to be doing real work — the model has presumably seen enough synthetic failure modes to generalize to real sensor corruption patterns, though the degree of that generalization will be scrutinized by reviewers.

## Why This Matters Beyond One Robot

The Unitree G1 is a reasonable test platform but not the commercial deployment target that most industry teams are building toward. The significance of CAP isn't the specific hardware — it's the architectural argument.

If the denoising world-model approach holds up, it suggests a path to significantly more robust perceptive policies without requiring sensor redundancy or hand-engineered failure detection. That's relevant to every team currently building perceptive locomotion stacks: Figure AI, Agility Robotics, and Boston Dynamics' Atlas program among them. It's also directly relevant to any team attempting [loco-manipulation](https://humanoidintel.ai/glossary/loco-manipulation) tasks where the robot must navigate to a manipulation site across unpredictable terrain.

The broader implication for the locomotion policy literature: the "clean-data assumption" may be a fundamental ceiling that the field needs to abandon systematically, not patch around with routing logic. CAP is one proposed solution. Expect competing approaches — possibly including learned confidence estimation, sensor fusion with IMU, or explicit occupancy-map reconstruction — to appear in response.

**One open question worth tracking:** how the denoising encoder performs when depth failure is total rather than partial. The paper's framing emphasizes partial corruption as the target regime. Complete sensor blackout — both eyes closed, so to speak — presumably still routes to proprioceptive-only behavior, but the boundary conditions aren't fully characterized in the abstract.

---

## Key Takeaways

- **CAP proposes a single-stage humanoid locomotion policy** that reconstructs clean depth from corrupted sensor input using a learned denoising world-model encoder — no hard switching between sub-policies.
- **A co-active proprioceptive variational encoder** runs in parallel at all times, providing depth-independent body-state information continuously rather than as a fallback.
- **Training uses a depth-noise curriculum plus feature dropout**, exposing the policy to the full spectrum of perception quality in one unified training run.
- **Validated on the Unitree G1** under intermittent occlusion, real-sensor corruption, and outdoor depth artifacts in both controlled trials and free-roaming deployments.
- **In simulation, CAP degrades more smoothly than binary-switching baselines** as perception worsens — a meaningful operational safety property beyond peak performance metrics.
- **The architectural bet is meaningful for the whole field**: abandoning the clean-data assumption in favor of learned denoising may generalize beyond one robot or one sensor type.

---

## Frequently Asked Questions

**What is the CAP locomotion system and how does it work?**
CAP (Continuously Adaptive Perception-Blind Humanoid Locomotion via Learned Denoising) is a single-stage locomotion policy for humanoid robots that uses a learned denoising encoder to reconstruct clean depth information from corrupted sensor input. It pairs this with a proprioceptive variational encoder that runs simultaneously, allowing the policy to handle any point on the spectrum from perfect depth sensing to complete depth failure without switching between separate sub-policies.

**Why do depth sensors fail on humanoid robots in real deployments?**
Depth sensors — structured light, time-of-flight, and stereo cameras — fail in real environments due to direct sunlight washing out IR signals, transparent or reflective surfaces producing phantom readings, and intermittent occlusion from the robot's own limbs or environmental obstacles. These failures are partial and intermittent, not binary, which is why the CAP authors argue that hard switching to blind control discards recoverable signal.

**What robot did the CAP researchers use for hardware validation?**
The researchers validated CAP on the Unitree G1 humanoid robot, running both controlled trials and indoor-outdoor deployments. Testing conditions included intermittent occlusion, real-sensor corruption, and outdoor depth artifacts.

**How is CAP different from previous perceptive-plus-blind locomotion approaches?**
Previous approaches either assume clean depth (perceptive-only) or switch to a separate blind sub-policy when perception degrades (routing/switching architectures). CAP uses a single policy trained across the entire perception-quality spectrum. The denoising encoder implicitly communicates perceptual confidence through its latent representation rather than requiring an explicit classifier to decide which sub-policy to activate.

**What are the implications for commercial humanoid deployments?**
If the denoising approach generalizes, it offers a path to more robust outdoor locomotion without requiring sensor redundancy hardware or hand-engineered failure detection logic. This is directly relevant to any team deploying humanoids in uncontrolled environments where lighting, weather, and occlusion conditions are variable — essentially every real-world deployment scenario beyond controlled factory floors.